Dither3D Shader — Audit & Fix Report

Date: 2026-05-15

Scope: Full codebase integrity review against original Dither3DInclude.cginc

1. BUG FIXES

1.1 — CRITICAL: Sky double exposure/offset

File: shaders/gbuffers_skybasic.fsh:37 Severity: High — visual blowout at any non-default exposure/offset

applyDither3DColor() already applies DITHER_EXPOSURE * color + DITHER_OFFSET internally (line 53 of dither3d_color.glsl). The sky shader was pre-applying the same transform:

// BEFORE (bug):
vec3 skyColor = clamp(glcolor.rgb * DITHER_EXPOSURE + DITHER_OFFSET, 0.0, 1.0);
vec3 result = applyDither3DColor(skyUV, screenPos, dx, dy, skyColor);
// Result: exposure and offset squared (over-exposed sky)

// AFTER (fix):
vec3 result = applyDither3DColor(skyUV, screenPos, dx, dy, glcolor.rgb);

All other gbuffer fragment shaders pass raw colors — only gbuffers_skybasic.fsh had this issue.

1.2 — MEDIUM: Invalid SVD frequency test

File: tests/test_math.py:15-30 Severity: Medium — test was validating a different algorithm than the shader uses

The Python compute_uv_frequency() computed sqrt(sqrt(lambda1 * lambda2)) — a geometric mean of eigenvalue square roots — instead of the SVD singular values the shader actually computes. The shader's algorithm:

float Q = dot(dx,dx) + dot(dy,dy);          // Frobenius norm squared
float R = dx.x*dy.y - dx.y*dy.x;            // determinant
float disc = sqrt(Q*Q - 4.0*R*R);
vec2 freq = sqrt(vec2(Q + disc, Q - disc) * 0.5);  // singular values (sigma_1, sigma_2)

The test now matches this exactly with three validated cases:

• Isotropic (uniform UV scale): both singular values equal scale ✓
• Anisotropic (stretched UV): singular values equal individual axis scales ✓
• Rotated (45°): singular values invariant under rotation ✓

Also fixed: Unicode characters (checkmarks, arrows) replaced with ASCII for Windows console compatibility.

2. CODE QUALITY IMPROVEMENTS

2.1 — Clarified dotRadius formula

File: shaders/lib/dither3d_core.glsl:98

// BEFORE (confusing):
float dotRadius = 0.25 / sqrt(activeDots * 0.25);

// AFTER (clear):
float dotRadius = 0.5 / sqrt(activeDots);

Both expressions are mathematically identical. The new form directly shows the relationship: radius scales as 0.5 / sqrt(N), giving 0.125 for 16 dots and 0.25 for 4 dots.

2.2 — PALETTE_COLOR_MATCH default changed to Color-Aware

Files: shaders/lib/dither3d_options.glsl:41, shaders/lib/dither3d_config.glsl:63

#define PALETTE_COLOR_MATCH 1  // was 0

Color-aware palette matching preserves hues by finding the two closest palette colors and dithering between them. This gives better results for colored palettes (CGA, Pico-8, Nord, Eldritch). The GAMEBOY profile explicitly overrides to 0 (luminance-based) for authentic retro look.

3. ARCHITECTURE VERIFICATION

3.1 — Include chain (correct)

Fragment shader
  -> dither3d_options.glsl     (defines all macros first)
  -> dither3d_color.glsl
       -> dither3d_core.glsl
            -> dither3d_config.glsl   (fallback defaults via #ifndef)
       -> dither3d_utils.glsl
       -> dither3d_palettes.glsl
            -> dither3d_utils.glsl    (guard prevents double include)

All 6 modules have unique, consistent include guards: | Module | Guard | |--------|-------| | dither3d_options.glsl | DITHER3D_OPTIONS_GLSL | | dither3d_config.glsl | DITHER3D_CONFIG_GLSL | | dither3d_core.glsl | DITHER3D_CORE_GLSL | | dither3d_utils.glsl | DITHER3D_UTILS_GLSL | | dither3d_palettes.glsl | DITHER3D_PALETTES_GLSL | | dither3d_color.glsl | DITHER3D_COLOR_GLSL |

3.2 — Vertex shaders (all 14 pairs verified)

All vertex shaders correctly compute and pass screenPos = gl_Position for radial compensation. World position and normal calculations are consistent across all geometry types:

• Terrain, Entities, Block, Hand, Water: triplanar UV with world normal
• Basic, Clouds, Weather, Armor Glint, Spider Eyes: simple world UV (no normal available)
• Sky Basic: cylindrical UV with view direction and alternate UV seam handling
• Sky Textured: texture atlas UV (scaled x4 for dithering)
• Beacon Beam: XZ plane with vertical offset

3.3 — Fragment shaders (all 14 pairs verified)

All fragment shaders follow the same pattern:

1. Include dither3d_options.glsl (first, defines macros)
2. Include dither3d_color.glsl (pulls in everything)
3. Sample textures, compute color
4. Call applyDither3DColorSimple() or applyDither3DColor()
5. Output to gl_FragData[0] with proper alpha preservation

3.4 — Shader properties (validated)

• shaders.properties: 14 RENDER_STYLE values, 8 profiles, proper slider ranges, nested submenu structure
• lang/en_US.lang: all options, values, profiles, and screen names have labels
• Custom palette sliders (colors 1-8, 24 RGB sliders) all wired through shaders.properties → dither3d_options.glsl → dither3d_config.glsl

3.5 — Palette system (validated)

• 9 built-in palettes: 1-Bit(2), GameBoy(4), CGA(4), VirtualBoy(4), Sepia(4), Nord(8), Solarized(8), Pico-8(16), Eldritch(16)
• 3 custom modes: 2-color, 4-color, 8-color via user-adjustable RGB sliders
• Dual matching: luminance-based (fast, crisp) and color-aware (preserves hues)
• All palettes sorted by luminance for correct dithering gradient
• getPaletteColor() centralizes all access with clamp() bounds safety

3.6 — CMYK halftone (validated)

• 4 plates with correct traditional screen angles (15°, 75°, 0°, 45°)
• Slight scale variation per plate breaks up moire patterns
• Round-trip conversion matches original HLSL (validated by test)

4. ORIGINAL PORT FIDELITY

Known differences (intentional adaptions for Minecraft):

• No 3D texture: replaced with procedural Bayer-pattern circular dots
• No brightness ramp texture: direct clamping used instead
• Contrast multiplier: 0.15 vs 0.1 (compensates for procedural pattern difference)
• contrastFade formula: 1/(1+c*0.5) vs 1.05/(1+c) (recalibrated for procedural dots)
• RENDER_STYLE 0 (RGB) uses small UV offsets instead of original's per-channel dithering

5. TEST RESULTS

python tests/test_math.py

======================================================================
Dither3D Mathematical Function Tests
======================================================================

--- SVD Frequency Computation (matches dither3d_core.glsl) ---
PASS Isotropic test:  maxFreq=2.000, minFreq=2.000, stretch=1.000
PASS Anisotropic test: maxFreq=4.000, minFreq=1.000, stretch=0.250
PASS Rotated test:     maxFreq=2.000, minFreq=2.000, stretch=1.000

--- CMYK Color Conversion ---
PASS Red, Green, Blue, Yellow, Magenta, Cyan, White, Black (8/8)

--- Bayer Matrix Generation ---
PASS Bayer 1x1, 2x2, 4x4, 8x8 (4/4)

======================================================================
All tests passed!
======================================================================

6. SUMMARY

Verdict: Codebase is stable, well-structured, and faithful to the original algorithm.